Sébastien Bubeck (born April 16, 1985) is a French-American machine learning researcher and the lead author of *Sparks of Artificial General Intelligence: Early experiments with GPT-4* (Bubeck et al., Microsoft Research, 2023). He was a vice president at Microsoft Research through early 2024 and has been a researcher at OpenAI since 2024. Earlier in his career he worked primarily on convex optimization, online algorithms, and the theory of machine learning, receiving best-paper awards at several theory conferences (Source: http://sbubeck.com/; Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck).
Background and education
Bubeck was born on April 16, 1985, and is described as a French-American computer scientist (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck). He studied at the École Normale Supérieure de Cachan (now ENS Paris-Saclay) and received his PhD from the Lille 1 University of Science and Technology (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck). His doctoral work was recognized with the Jacques Neveu prize for the best French PhD in probability and statistics, and with prizes in France's AfIA AI-thesis awards and the Gilles Kahn prize for a French PhD in computer science (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck).
Background and roles
Before joining Microsoft Research, Bubeck spent three years as an assistant professor at Princeton University, in the Department of Operations Research and Financial Engineering (Source: http://sbubeck.com/; Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck). He had also been a researcher at the University of California, Berkeley (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck).
He then spent ten years at Microsoft Research, first joining its Theory Group (Source: http://sbubeck.com/). By his own account and Wikipedia's, he rose to vice president of applied research and distinguished scientist and led the Machine Learning Foundations group at Microsoft Research Redmond (Source: http://sbubeck.com/; Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck).
In October 2024, Bubeck left Microsoft to join OpenAI, a move reported by The Information and by Bloomberg (Source: https://www.theinformation.com/briefings/microsoft-ai-researcher-sebastien-bubeck-to-join-openai; Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck). The move placed his capability-evaluation work inside the lab whose model he had previously studied from outside.
Research
In the first part of his career Bubeck worked mainly on convex optimization, online algorithms, and adversarial robustness in machine learning (Source: http://sbubeck.com/). His theoretical contributions include minimax rates for multi-armed bandits and linear bandits, an optimal algorithm for bandit convex optimization, and work on the k-server and metrical task systems problems (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck). He is the author of the monograph Convex Optimization: Algorithms and Complexity (2015), published in Foundations and Trends in Machine Learning (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck).
With Mark Sellke, Bubeck introduced and proved what they termed a "universal law of robustness," a result relating the number of parameters in a neural network to its smoothness, which holds that smooth interpolation of data requires substantially more parameters than mere fitting (Source: https://arxiv.org/abs/2105.12806). The paper, "A Universal Law of Robustness via Isoperimetry," received an Outstanding Paper Award at NeurIPS 2021 (Source: https://nips.cc/virtual/2021/awards_detail). Across his theory work he received best-paper or best-student-paper awards at COLT (2009, 2016), NeurIPS (2018, 2021), ALT (2018, 2023), and STOC (2023), and a 2015 Alfred P. Sloan Research Fellowship in computer science (Source: http://sbubeck.com/; Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck). He has served on the editorial board of the Journal of the ACM and as program-committee chair of the 2018 Conference on Learning Theory (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck).
Bubeck describes his later work as shifting toward understanding how intelligence emerges in large language models and how that understanding can be used to improve them, an approach he and his collaborators call "Physics of AGI," which examines how a model's parts (parameters, neurons, layers, data curriculum) combine to produce its behavior (Source: http://sbubeck.com/).
Small language models (Phi)
At Microsoft Research, Bubeck led work on a family of small language models released under the name Phi. The 2023 paper "Textbooks Are All You Need," with Bubeck among the authors, introduced phi-1, a code-generation model that the authors reported reached high performance at a size substantially smaller than competing models by training on data the authors describe as "textbook-quality" (Source: https://arxiv.org/abs/2306.11644). Microsoft followed with Phi-2, a 2.7-billion-parameter model the company said matched or exceeded larger models on reasoning and language-understanding benchmarks (Source: https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/). This line of work received coverage in outlets including the New York Times, the Wall Street Journal, Wired, and Scientific American, several of which framed it around the question of whether larger models are always better (Source: http://sbubeck.com/).
Sparks of AGI
At Microsoft Research, Bubeck led the 150+ page report that argued an early version of GPT-4 should be viewed as an early, incomplete form of AGI. The report established qualitative breadth testing across domains — math, coding, vision, medicine, law, and psychology — as a capability-reporting style, and it coincided with wider use of the term "AGI" in peer-reviewed and arXiv work. Its capability claims served as an anchor that subsequent AGI-timeline essays, including Aschenbrenner's *Situational Awareness*, extrapolate from.
The report has been both influential and contested. Skeptics note its qualitative, hand-picked methodology and Microsoft's institutional position as an OpenAI investor. Wikipedia records that the paper drew wide interest and debate in both the scientific community and the popular press, with coverage in the New York Times, Wired, and elsewhere (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck). In a related 2023 paper in the New England Journal of Medicine, Bubeck and co-authors Peter Lee and Joseph Petro examined the benefits, limits, and risks of using GPT-4 as an AI chatbot for medicine (Source: https://en.wikipedia.org/wiki/S%C3%A9bastien_Bubeck).
Relationships
- related: Sparks of AGI (lead author)
- related: OpenAI (current), Microsoft Research (former)
- supports: General-Purpose AI, AGI Timelines
- related: AI Benchmarks and Evaluation — Sparks' qualitative method itself a subject of benchmark debate